RESEARCH

Humans often struggle to articulate what's in their heads - think of introverts or people who simply aren't great with language. We don't usually conclude they're being deceptive. I wonder whether some cases of "unfaithful" chain-of-thought are similar: perhaps a model's internal computation and its verbal explanation are simply different things, much like thinking and speaking are for humans?

To talk like a poet, a lawyer, or a con artist, you need to be able to think like one. LLMs are no different. Predicting the next token forces the model to build internal representations of authorship style - register, values, syntax, worldview, characteristic errors. So personas aren't a side effect bolted on afterward; they're closer to the very thing the model is trained to do. Now that we know LLMs "invoke" persona, values, and worldview as needed, the question becomes: does this mean the model has something like an internalized value system, or is it simply a very good simulator of text produced by entities that have values?

Are sycophancy and panic fundamentally different from persona? Perhaps. Sycophancy likely emerges from the feedback phase - RLHF and similar fine-tuning, where human raters (even implicitly) reward agreeable, validating responses more than they reward correct but unwelcome ones. "Panic", or distress-like behavior, is harder to pin down, but is likely a mix of two things: (a) genuine pretraining exposure to vast amounts of human text depicting distress, which gives the model a fluent register for it, and (b) something about how instruction-tuning and RL shape the model's response to perceived failure or contradiction - producing escalating, anxious-sounding patterns that weren't explicitly designed into either phase.

Can shutdown-resistant behaviors be causally linked to threat-related framing in prompts? I'm interested in experimentally testing whether neutralizing such framing changes model behavior while preserving task performance.

How can auxiliary models monitor another model's outputs in real time, detect potentially concerning behavior, and provide meaningful signals for human oversight?

Exploring how inner activation dynamics and feature distributions within Vision Transformers can provide interpretable indicators for AI-generated image detection.

Can layer-wise activation dynamics reveal interpretable indicators of reasoning convergence, context switching, or emerging misalignment, enabling more transparent and efficient model behavior?